A Method and System for Monitoring Electricity Trading Risks Based on Multi-Source Signaling Mechanism and Sparse Hybrid Expert Model

Abstract

This invention proposes a method and system for monitoring electricity trading risks based on a multi-source signal mechanism and a sparse hybrid expert model. The method includes: collecting historical day-ahead electricity market price data and raw data on influencing factors; performing data preprocessing to construct a multi-source fusion signal from price spread signals and bidding space signals; using a random forest algorithm for feature selection; constructing a set of similar day data based on the bidding space signal and incorporating it into the training set; using a sparse hybrid expert model as the model framework, with the price spread signal and bidding space signal as targets, training different prediction models; and searching for an optimal declaration curve based on the Northern Eagle optimization search algorithm by predicting the price spread signal and bidding space signal, using the short-term power prediction curve as a benchmark. This invention effectively solves the problem of insufficient model generalization ability, realizes risk monitoring in electricity spot trading, and improves the accuracy of electricity spot trading risk assessment while enhancing the stability of the power system.

Description

Technical Field

[0001] This invention belongs to the field of new energy power technology, and in particular relates to a power trading risk monitoring method and system based on a multi-source signal mechanism and a sparse hybrid expert model. Background Technology

[0002] With the continuous development of the electricity market, the importance of electricity trading in power system operation, electricity market transactions, and energy management is becoming increasingly prominent. Accurate electricity trading can effectively provide important support for the stable operation of the electricity market and the optimization of resource allocation. However, the electricity market is constantly affected by a variety of complex factors, including supply and demand, fuel prices, weather conditions, and fluctuations in renewable energy output. The interaction between these factors makes electricity trading highly nonlinear and uncertain, which directly exacerbates the power imbalance in the system. Traditional methods are insufficient to meet the growing demand for accuracy in monitoring electricity trading risks.

[0003] In recent years, with the rapid development of artificial intelligence technology, especially the widespread application of deep learning and big data technologies, data-driven methods for power trading forecasting and monitoring have gradually become a research hotspot. However, forecasting and monitoring models based on a single data source often struggle to fully capture the complex changing patterns of power trading, resulting in limited forecasting and monitoring accuracy.

[0004] From a technical perspective, the shortcomings of existing power trading risk monitoring can be summarized as follows: (1) Data dependence and quality defects, for example, the price spread signal is affected by weather, sudden events, etc., the data fluctuates greatly and contains noise. Traditional filtering methods (such as wavelet denoising) may lose key information, and the power trading risk caused by the fluctuation of the price spread signal is difficult to predict accurately.

[0005] (2) Insufficient integration of multi-source data. Traditional methods use a single data source or simply splice multi-source data, resulting in low utilization of market microstructure information.

[0006] (3) The model has insufficient generalization ability and there is a risk of overfitting. Deep learning models are prone to overfitting when training data is limited, and it is difficult to adapt to new markets or extreme scenarios.

[0007] (4) Poor dynamic adaptability. Market rule changes (such as the introduction of carbon emission trading) or sudden events require the model to learn online, but most static models need to be retrained frequently.

[0008] (5) Optimization layer defects, limitations of search algorithms, slow convergence speed and unstable solution quality of traditional optimization methods (such as genetic algorithms) on problems with 96 time points.

[0009] (6) Limitations of complex nonlinear relationship modeling: Traditional methods (such as SVM and random forest) are insufficient in capturing the high-order nonlinear relationship between power trading risk and multiple factors (such as wind power output and fuel price).

[0010] These shortcomings reflect that power trading risk monitoring still needs to make breakthroughs in the collaborative optimization of data, models, computing, and commercial implementation in terms of technical logic.

[0011] Multi-source data fusion technology, by integrating data from different sources (such as historical electricity price data, meteorological data, load data, thermal power output data, wind and solar power output data, etc.), can more comprehensively reflect the multi-dimensional factors affecting electricity trading risks, thereby improving the accuracy and robustness of prediction and monitoring models. Furthermore, large-scale models have achieved significant results in natural language processing, image recognition, and other fields, and their powerful feature extraction and modeling capabilities provide a new technical approach for electricity trading risk monitoring.

[0012] How to combine multi-source data fusion with large-scale models to fully leverage the advantages of large-scale models in processing high-dimensional and nonlinear data, and further improve the accuracy of power trading risk monitoring, has become an urgent problem to be solved. Summary of the Invention

[0013] This invention proposes a method and system for monitoring power trading risks based on a multi-source signal mechanism and a sparse hybrid expert model. This effectively solves the problem of insufficient model generalization ability, realizes risk monitoring of power spot trading, and improves the accuracy of power spot trading risk assessment while enhancing the stability of the power system.

[0014] To achieve the above objectives, the technical solution of the present invention is implemented as follows: A power trading risk monitoring method based on a multi-source signaling mechanism and a sparse hybrid expert model includes: S1. Collect historical day-ahead electricity market price data and raw data of influencing factors, perform data preprocessing, and construct a dataset; based on the dataset, construct price difference signals and bidding space signals, and combine dynamic weighted fusion algorithm and rolling window smoothing technology to generate multi-source fusion signals; S2. The random forest algorithm is used to calculate the importance weights of each influencing factor and the multi-source fusion signal on the price difference signal. Feature screening is performed, and the top_p features with the highest importance weights are selected to construct the training set and the test set. S3. Construct a similar day data set based on the bidding space signal, and incorporate the similar day data set into the training set; S4. The SparseMoE architecture of the sparse hybrid expert model is used as the model framework, and the feedforward neural network model is used as the basic expert model. Based on the training set data, different prediction models are trained with the price difference signal and the bidding space signal as the targets. S5. Based on the predicted price difference signal and bidding space signal obtained from the prediction model, a profit objective function is constructed. The assessment mechanism function is used as a penalty term, and the multi-source fusion signal is used as a condition. The original short-term power prediction is used as the strategy input parameter. The optimal application curve for the prediction day is dynamically searched through the Northern Eagle optimization search algorithm. Risk monitoring is achieved by comparing the consistency between the user application curve and the optimal application curve.

[0015] Furthermore, the preprocessing method for the raw data of influencing factors in step S1 includes: S101. Use the density-based DBSCAN clustering algorithm to perform outlier detection and analysis, and set outliers to null. S102. A three-stage hybrid filling strategy is adopted to fill null values ​​in the data.

[0016] Furthermore, the method for generating the multi-source fused signal in step S1 includes: S111, Constructing price spread signals and bidding space signals; S112. The correlation coefficient reflects the synergistic relationship between the price spread signal and the bidding space signal. Calculate the benchmark weights; ;in Represents the benchmark weight; This represents the weight sensitivity coefficient, with a value range of [0.1, 0.3]. S113. Calculate the dynamic weight of the bidding space signal using the benchmark weight; ; represent t Dynamic weighting of the bidding space signal at any given time; For bidding space signals in the window L within the standard deviation, For bidding space signals in the window L The mean within; S114. Signal fusion is performed based on dynamic weights to generate multi-source fused signals; ; Represents a multi-source fused signal; represent t Real-time bidding space signal; represent t Time-based price spread signals.

[0017] Furthermore, step S2 includes: The random forest algorithm is used to select the top_p features with the highest importance weights. ; in, It is a tree tree All features used in f The set of nodes to be split, features f These are all feature columns except the target value; Impurity(s) It is a node s impurity, , , These are the parent node, left child node, and right child node, respectively. It is a tree tree Chinese characteristics f The weight.

[0018] Furthermore, step S3 includes: Based on the bidding space signal of the prediction date, the data of the T dates closest to the bidding space signal are selected from the historical data through Pearson correlation coefficient calculation to construct a similar day data set for the prediction date, and the similar day data set is included in the training set.

[0019] In another aspect, this invention proposes a power trading risk monitoring system based on a multi-source signaling mechanism and a sparse hybrid expert model, comprising: Preprocessing and multi-source fusion signal module: Collects historical day-ahead electricity market price data and raw data of influencing factors, performs data preprocessing, and constructs a dataset; based on the dataset, constructs price difference signals and bidding space signals, and generates multi-source fusion signals by combining dynamic weighted fusion algorithm and rolling window smoothing technology; Weighting module: The random forest algorithm is used to calculate the importance weight of each influencing factor and the multi-source fusion signal on the price difference signal. Feature selection is performed to select the top_p features with the highest importance weights to construct the training set and test set. Similar Day Module: Construct a similar day dataset based on the bidding space signal, and incorporate the similar day dataset into the training set; Prediction Model Module: The SparseMoE architecture of the sparse hybrid expert model is used as the model framework, and the feedforward neural network model is used as the basic expert model. Based on the training set data, different prediction models are trained with the price spread signal and the bidding space signal as the targets. Risk monitoring module: Based on the price difference signal and bidding space signal obtained from the prediction model, a profit objective function is constructed. The assessment mechanism function is used as a penalty term, and multi-source fusion signals are used as conditions. The original short-term power prediction is used as the strategy input parameter. The optimal application curve for the prediction day is dynamically searched through the Northern Eagle optimization search algorithm. Risk monitoring is achieved by comparing the consistency between the user's application curve and the optimal application curve.

[0020] Furthermore, the preprocessing and multi-source fusion signal module includes: Outlier handling unit: Employs density-based DBSCAN clustering algorithm for outlier detection and analysis, and sets outliers to null. Null value filling unit: A three-stage hybrid filling strategy is used to fill null values ​​in the data.

[0021] Furthermore, the preprocessing and multi-source fusion signal module includes: Signal construction unit: Constructs price spread signals and bidding space signals; Benchmark weighting unit: The correlation coefficient reflects the synergistic relationship between the price spread signal and the bidding space signal. Calculate the benchmark weights; ;in Represents the benchmark weight; This represents the weight sensitivity coefficient, with a value range of [0.1, 0.3]. Dynamic weighting unit: Calculates the dynamic weight of the bidding space signal using the benchmark weight; ; represent t Dynamic weighting of the bidding space signal at any given time; For bidding space signals in the window L within the standard deviation, For bidding space signals in the window L The mean within; Fusion Unit: Performs signal fusion based on dynamic weights to generate multi-source fused signals; ; Represents a multi-source fused signal; represent t Real-time bidding space signal; represent t Time-based price spread signals.

[0022] Furthermore, the weighting module includes: The random forest algorithm is used to select the top_p features with the highest importance weights. ; in, It is a tree tree All features used in f The set of nodes to be split, features f These are all feature columns except the target value; Impurity(s) It is a node s impurity, , , These are the parent node, left child node, and right child node, respectively. It is a tree tree Chinese characteristics f The weight.

[0023] Furthermore, the similar day module includes: Based on the bidding space signal of the prediction date, the data of the T dates closest to the bidding space signal are selected from the historical data through Pearson correlation coefficient calculation to construct a similar day data set for the prediction date, and the similar day data set is included in the training set.

[0024] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention effectively solves the problem of insufficient model generalization ability, realizes risk monitoring of electricity spot trading, and improves the accuracy of electricity spot trading risk assessment while improving the stability of the power system.

[0025] 2. This invention proposes a multi-source data fusion architecture. By integrating multiple data sources, it proposes a dual-drive mode of "physical characteristics + market signals" and a dynamic weighted fusion mechanism of "price difference signals + bidding space signals", which solves the problem of insufficient utilization of micro-structure information in the power market in traditional methods.

[0026] 3. In signal prediction, this invention adopts a sparse hybrid expert architecture based on a feedforward neural network model with dynamic routing gating. Compared with traditional time series prediction large models, it achieves higher prediction accuracy with fewer parameters, effectively solving the limitations of complex nonlinear relationship modeling and the problem that traditional methods are insufficient in capturing high-order nonlinear relationships.

[0027] 4. In terms of risk monitoring, this invention uses the Northern Eagle Optimization Algorithm as the search algorithm, and dynamically searches for the optimal solution through two-stage hunting behavior. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the overall process of Embodiment 1 of the present invention; Detailed Implementation

[0029] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0030] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0031] Example 1: This embodiment proposes a power trading risk monitoring method based on a multi-source signal mechanism and a sparse hybrid expert model. The core design follows this logic: data preprocessing → constructing multi-source signals → data filtering → constructing a similar daily data set → constructing a prediction model → model training → prediction output → constructing an assessment mechanism and optimization objective → introducing the Northern Eagle optimization search algorithm → outputting the optimal declaration curve → comparison and early warning. Figure 1 The diagram shown is the overall flowchart, including: 1. Data collection and preprocessing.

[0032] (1) Collect historical day-ahead electricity market supply and demand data and raw data of its influencing factors, conduct preliminary screening based on historical price difference signals and raw data of its influencing factors, and determine the content of multi-source information data, including photovoltaic power generation output, wind power generation output, power load, hydropower including pumped storage power generation output, total new energy output, inter-provincial interconnection lines, thermal power output, total power generation, minimum thermal power, and maximum thermal power.

[0033] (2) The density-based DBSCAN clustering algorithm is used for outlier detection and analysis. By defining two core parameters, neighborhood radius (eps) and minimum number of samples (min_samples), noise points and outliers in the data distribution are automatically identified. For sample points that are identified as outliers, the outlier is set to empty, and the outlier is recorded for subsequent analysis.

[0034] (3) A three-stage hybrid filling strategy is adopted. For a single null value in a column of feature data, forward filling (ffill) is used to inherit the valid observations from the previous stage, and backward filling (bfill) is used to supplement the subsequent valid data. The sum and average of the two filling results are used to obtain the final filled value. Secondly, for a column of feature data with consecutive null values, forward filling and backward filling are used for the first and last data respectively. In addition, zero-value filling is performed for consecutive missing segments in the middle. This method not only maintains the temporal continuity of the data, but also avoids the problem of bias accumulation caused by a single filling method, and ensures numerical stability in extreme cases through zero filling.

[0035] The dataset is constructed after preprocessing and used for training subsequent models.

[0036] 2. Construct a multi-source fusion signal.

[0037] This step uses price spread signals and bidding space signals as the core logic for power trading, and combines a dynamic weighted fusion algorithm with rolling window smoothing technology to generate multi-source fused signals. Ultimately, a multi-source signal mechanism is formed for risk early warning.

[0038] The core power trading logic combines a dynamic weighted fusion algorithm with rolling window smoothing technology to generate multi-source fusion signals. These signals are then integrated from price spread signals, bidding space signals, and multi-source fusion signals, using a voting system to guide strategy direction in parallel. Ultimately, this multi-source signal mechanism guides strategy direction and indirectly contributes to risk warning.

[0039] (1) Construct price spread signals and bidding space signals; The method for constructing the price spread signal is as follows: Price spread = Day-ahead electricity price - Real-time electricity price; if the day-ahead electricity price is greater than the real-time electricity price, the price spread signal is 1; if the day-ahead electricity price is less than or equal to the real-time electricity price, the price spread signal is -1. The method for constructing the bidding space signal includes: Bidding Space = Electricity Load - Total Output of New Energy Sources - External Power Supply Plan. After sorting the bidding space signals, the median of the bidding space is found through quartile calculation. For example, if the median is 1W, then 1W is used as the boundary; if it is greater than 1W, the bidding space signal is 1; if it is less than or equal to 1W, the bidding space signal is -1. The external power supply plan is derived from data published by the power trading center, which can be obtained from power trading centers in different provinces.

[0040] (2) Calculate the benchmark weights; Based on the established price spread and bidding space signals, the correlation coefficient reflects the synergistic relationship between the price spread and bidding space signals in real time. A volatility adjustment term automatically amplifies the weight of the price spread signal during periods of high volatility.

[0041] The formula for calculating the benchmark weight is: ; in Represents the benchmark weight; For price spread signals and bidding space signals in the window L The correlation coefficient (default 96 time points). As a volatility adjustment term, when the spread signal and the bid-ask space signal are positively correlated... >0.5, enhance benchmark weight When the price spread signal and the bidding space signal are negatively correlated ( If the value is less than 0.5, then the benchmark weight is reduced. . Represents the weight sensitivity coefficient, controlling the correlation coefficient. For the benchmark weight The influence strength, with a value range of [0.1, 0.3], maintains the benchmark weight around 0.5 to avoid drastic fluctuations in weight; the default value is 0.3. In this formula, the correlation coefficient between the price difference signal and the bidding space signal... When = 0.5, the benchmark weight It equals 0.5.

[0042] Furthermore, a weight truncation mechanism is employed to prevent signal instability under extreme market conditions. This weight truncation mechanism refers to truncating the base weights that are less than 0. Set to 0 for baseline weights greater than 1. Set to 1 to ensure that the baseline weight remains within the [0,1] range and prevent weights from going out of bounds.

[0043] (3) Calculate the dynamic weights; The dynamic weight of the bidding space signal is calculated using a benchmark weight; the calculation formula is as follows: ; represent t Dynamic weighting of the bidding space signal at any given time; For bidding space signals in the window L within the standard deviation, For bidding space signals in the window L The mean within; The dynamic weights are calculated based on the benchmark weights. During periods of high volatility, the volatility adjustment term amplifies the benchmark weights, thus automatically increasing the dynamic weights; and through... clip The function ensures that the dynamic weights remain within the [0,1] interval to prevent overfitting.

[0044] (4) Signal fusion is performed based on dynamic weights to generate multi-source fused signals; ; Represents a multi-source fused signal; represent t Real-time bidding space signal; represent t Time-based price spread signals.

[0045] The multi-source fusion signal is used to determine the optimal application curve.

[0046] (5) Standardize the dataset to eliminate the influence of units on model training.

[0047] (6) Split the dataset: use the historical data from the three months prior to the prediction date for training, and use the data from the prediction date itself for prediction.

[0048] 3. Data filtering.

[0049] The random forest algorithm is used to calculate the importance of each feature (all feature columns except the target value) to the price spread signal, and different features are assigned different importance weights. The target value is the price spread signal and the bidding space signal. All feature columns other than the target value include the influencing factors of the price spread signal collected in step 1, and the multi-source fusion signal constructed in step 2. .

[0050] The formula for data filtering in random forest is as follows: ; in, It is a tree tree All features used in f The set of nodes to be split, features f These are all the feature columns except the target value; Impurity(s) It is a node s impurity, , , These are the parent node, left child node, and right child node, respectively. It is a tree tree Chinese characteristics f Importance weights.

[0051] The formula for calculating the impurity is as follows: The impurity of a node is represented by information entropy, which measures the degree of disorder of sample categories in a node. The smaller the entropy, the higher the purity.

[0052] ; in The information entropy of a node is represented by K, where K is the total number of sample classes and k represents the k-th class sample. It is the proportion of the k-th class of samples, when When =0, define It is 0.

[0053] After calculating the importance weight of each feature on the price spread signal, the top_p features with the highest importance weights are selected to construct the training set and the test set.

[0054] 4. Construct a set of similar daily data.

[0055] (1) Based on the bidding space signal of the forecast date (bidding space of the forecast date = load of the forecast date - total output of new energy on the forecast date - external power plan on the forecast date), the data of the T dates closest to the bidding space signal are selected from the historical data through the Pearson correlation coefficient calculation to construct a similar day data set of the forecast date.

[0056] The Pearson correlation coefficient is used to calculate similarity using the following formula: ; in, It's similarity. It is the covariance of variables X and Y. It is the standard deviation of X. X is the standard deviation of Y. X and Y are data (feature columns) of different influencing factors in the dataset, and a set of similar days is constructed based on similarity.

[0057] (2) After the similar day data set is constructed, the similar day data set is included in the training set to participate in the model training.

[0058] In this embodiment, since the historical data of the three months (90 days) before the prediction date were previously split into a training set, the historical data from 90 days ago can also be included in the training set based on similarity through the above steps, which is equivalent to expanding the training set.

[0059] 5. Build a prediction model.

[0060] The SparseMoE (Sparse Mixture of Experts) architecture is used as the prediction model framework. Based on the training set data, prediction models are trained with price spread signals and bidding space signals as targets respectively. Based on the training set data and the weights, different prediction models are trained with price spread signals and bidding space signals as targets, including a price spread signal prediction model and a bidding space signal prediction model.

[0061] The SparseMoE (SparseMixtureofExperts) architecture is a neural network architecture that expands the model capacity through sparse activation. The core idea is to distribute computation across multiple expert networks, each expert network focusing on processing a specific data pattern, but each input only activates a small number of experts. Through a learnable gating network, the input is dynamically distributed to the outputs of the top k most relevant experts for integration, thereby maintaining computational efficiency while significantly increasing the number of parameters.

[0062] The prediction model uses a feedforward neural network model as the base expert model (expert network), employing multiple feedforward neural network models. The gating mechanism uses a dynamic routing iterative optimization approach, gradually adjusting the routing weight allocation between the input and the expert network. Dynamic routing, the process of assigning input data to different expert networks, is the core mechanism of the MoE architecture. It mainly consists of three steps: (1) Routing function: Receives the input representation and outputs the selection probability distribution for each expert network; as shown in the following formula: ; in, It is the routing weight matrix ( (where d is the number of expert networks and d is the vector embedding dimension). is an optional bias term; g is the original routing score, and the output of each expert network in dynamic routing selection is determined according to the score of g.

[0063] (2) Top-k selection: Selecting an expert network based on the probability distribution output by the routing function; ; in TopKIndices Indicating the selection of expert networks, p It is a probability distribution vector, representing the probability that the current input is assigned to each expert network; k This is the number of expert networks selected each time (hyperparameter), usually taken as 1 or 2; argtopk Used to obtain the probability distribution vector p Center front k The probability of a maximum or minimum value The probability of unselected expert networks is set to 0 to ensure sparsity.

[0064] (3) Weighted combination: The selected expert network processes the input and produces the output. The final output is the weighted sum of the outputs of each expert network. The routing weight is determined by the probability of the expert network. For the selected k expert networks, calculate their outputs and sum them using weighted averages: ; in, It is the output of the i-th expert network. It represents the probability that the current input is assigned to the i-th expert network.

[0065] Finally, through model training and evaluation, the outputs of the best Top-k experts are selected and superimposed to obtain the price spread signal prediction model and the bidding space signal prediction model, respectively. The price spread signal and the bidding space signal are then predicted and used as the final signal prediction results for the next step of risk monitoring and early warning.

[0066] 6. Risk monitoring and early warning.

[0067] (1) Construct a profit objective function, with the assessment mechanism function as the penalty term, the multi-source fusion signal as the condition, and the original short-term power prediction as the input parameter; use the Northern Eagle optimization search algorithm in step (2) below to search for the optimal declared electricity (forming the optimal declaration curve) and confidence space for the next 96 time points. The next 96 time points are the 96 time points with an interval of 15 minutes (min) between the prediction days.

[0068] The assessment mechanism function is determined by the rules of different provinces, and different provinces have different assessment mechanisms for electricity energy in the electricity trading market.

[0069] The formula for calculating the objective function is summarized as follows: (Optimal reporting curve - original short-term forecast curve) * (forecast day-ahead electricity price - forecast real-time electricity price) / 4; The original short-term prediction curve is obtained based on the original short-term power prediction of the input parameters. The original short-term power prediction is a conventional technique in this field and is obtained through a short-term power prediction model.

[0070] First, the strategy direction is guided by a voting system that integrates price spread signals, bidding space signals, and multi-source fusion signals. Among them, the price spread signals and bidding space signals are predicted by the prediction model in step 5.

[0071] For example, the price spread signal essentially compares the predicted day-ahead electricity price with the predicted real-time electricity price. Its impact on the strategy direction is as follows: if the predicted day-ahead electricity price is greater than the predicted real-time electricity price, the signal is set to 1, and the strategy direction is raised; otherwise, it is set to 0, and the strategy direction is lowered.

[0072] For example, the bidding space signal uses 1W as the dividing line. If it is greater than 1W, the strategy direction is raised; if it is less than or equal to 1W, the strategy direction is lowered. For example, if the fusion signal is greater than 0.5, the strategy direction is raised; if it is less than or equal to 0.5, the strategy direction is lowered. Finally, a voting process is conducted to determine whether the strategy direction is lowered or raised at each time point.

[0073] After deciding whether to raise or lower the power volume at each time point, and combining the assessment rules of different provinces, the Northern Eagle optimization algorithm is used to input the boundary conditions of the declared power volume at each time point [0, installed capacity]. Finally, in the specified direction, the assessment mechanism function is used as the penalty item, and the revenue situation under different declared power volumes is continuously calculated through the revenue objective function. Finally, the declared power volume corresponding to the optimal revenue situation at each time point is obtained.

[0074] The confidence space is based on the optimal reported power volume at each time point, plus or minus 5% * installed capacity, to enhance the generalization performance of the algorithm and provide some room for manual adjustment based on actual conditions.

[0075] (2) By using the Northern Eagle optimization search algorithm, the optimal declaration curve and confidence space for the next 96 time points are dynamically searched.

[0076] (2.1) Input optimization problem; The optimization problem primarily aims to establish an optimal reporting curve, which can both benefit the generation side and prevent market disruptions for the dispatching side. The goal is to prevent generation sides from unilaterally modifying their short-term forecasting curves in the electricity spot market, leading to significant losses. Alternatively, it aims to prevent generation sides from maliciously and drastically modifying their reporting curves for arbitrage, causing grid connection difficulties and disrupting the electricity market order.

[0077] (2.2) Set the number of iterations T and the population size N; The population size N represents the number of eagles. It can be understood as N representing that there are N different optimization solutions in each iteration, which are compared at the same time.

[0078] In this embodiment, the population size (number of eagles) refers to the number of solutions. For example, if N is set to 100, each iteration will generate 100 reporting curves. Then, the optimal reporting curve among the 100 is calculated and recorded. After recording the value of this optimal reporting curve and the optimal solution, the next iteration with N=100 will begin.

[0079] (2.3) Randomly initialize the population locations; The population location represents the location of each eagle, which in this embodiment is the reported electricity volume at each moment of the reporting curve, with a value of [0, installed capacity]. (2.4) Iteration t=1:T; (2.5) Iterate i=1:N; Phase 1: Identifying Prey (Exploration Phase); (2.7) Randomly select prey; In this embodiment, the prey representative considers the benefits after the assessment.

[0080] (2.8) Iterate j=1:m; Where m represents the number of times each eagle moves its position, which in this embodiment is the number of changes in the optimal reporting curve.

[0081] (2.9) Randomly update the current position of the eagle in the j-th dimension; The j-dimensional position refers to the current reported power consumption (eagle position) when iterating to the j-th iteration.

[0082] (2.10) End the iteration j=1:m; (2.11) Update the position of the i-th eagle; Phase 2: Following and chasing prey (exploration phase); (2.13) Update parameter R; The parameter R is a perturbation factor for updating the position of the eagle, which increases the ability of each eagle to jump out of the local optimum.

[0083] (2.14) Iterate j=1:m; (2.15) Calculate the current position information of the eagle in the j-th dimension; (2.16) End the iteration j=1:m; (2.17) Update the position of the i-th eagle; (2.18) End the iteration i=1:N; (2.19) End the iteration t=1:T; (2.20) Output the optimal solution to the problem.

[0084] The optimal solution to the problem is an optimal reporting curve. The confidence space is based on the optimal reported electricity volume at each time point, plus or minus 5% * installed capacity to enhance the generalization performance of the algorithm and provide some room for manual adjustment based on actual conditions.

[0085] (3) Finally, intelligent high and medium risk warnings are implemented by comparing the consistency between the user's declaration curve and the optimal declaration curve.

[0086] If the former and the latter show different positive and negative values, it indicates that the spot market reporting strategy is going in the opposite direction, and there is a high probability that the power generation side will suffer large losses, requiring a high level of warning.

[0087] If the former and the latter show the same positive and negative values, but a comparison of the user's declaration at each time point reveals that the declaration point is greater than the confidence space given by this method at each time point, it indicates that there may be a high degree of malicious arbitrage, and it may trigger a major assessment mechanism, resulting in a moderate warning.

[0088] This embodiment effectively solves the problem of insufficient model generalization ability, realizes risk monitoring of electricity spot trading, and improves the accuracy of electricity spot trading risk assessment while enhancing the stability of the power system.

[0089] Example 2: This embodiment proposes a power trading risk monitoring system based on a multi-source signaling mechanism and a sparse hybrid expert model, including: Preprocessing and multi-source fusion signal module: Collects historical day-ahead electricity market price data and raw data of influencing factors, performs data preprocessing, and constructs a dataset; based on the dataset, constructs price difference signals and bidding space signals, and generates multi-source fusion signals by combining dynamic weighted fusion algorithm and rolling window smoothing technology; Weighting module: The random forest algorithm is used to calculate the importance weight of each influencing factor and the multi-source fusion signal on the price difference signal. Feature selection is performed to select the top_p features with the highest importance weights to construct the training set and test set. Similar Day Module: Construct a similar day dataset based on the bidding space signal, and incorporate the similar day dataset into the training set; Prediction Model Module: The SparseMoE architecture of the sparse hybrid expert model is used as the model framework, and the feedforward neural network model is used as the basic expert model. Based on the training set data, different prediction models are trained with the price spread signal and the bidding space signal as the targets. Risk monitoring module: Based on the price difference signal and bidding space signal obtained from the prediction model, a profit objective function is constructed. The assessment mechanism function is used as a penalty term, and multi-source fusion signals are used as conditions. The original short-term power prediction is used as the strategy input parameter. The optimal application curve for the prediction day is dynamically searched through the Northern Eagle optimization search algorithm. Risk monitoring is achieved by comparing the consistency between the user's application curve and the optimal application curve.

[0090] Furthermore, the preprocessing and multi-source fusion signal module includes: Outlier handling unit: Employs density-based DBSCAN clustering algorithm for outlier detection and analysis, and sets outliers to null. Null value filling unit: A three-stage hybrid filling strategy is used to fill null values ​​in the data.

[0091] Furthermore, the preprocessing and multi-source fusion signal module includes: Signal construction unit: Constructs price spread signals and bidding space signals; Benchmark weighting unit: The correlation coefficient reflects the synergistic relationship between the price spread signal and the bidding space signal. Calculate the benchmark weights; ;in Represents the benchmark weight; This represents the weight sensitivity coefficient, with a value range of [0.1, 0.3]. Dynamic weighting unit: Calculates the dynamic weight of the bidding space signal using the benchmark weight; ; represent t Dynamic weighting of the bidding space signal at any given time; For bidding space signals in the window L within the standard deviation, For bidding space signals in the window L The mean within; Fusion Unit: Performs signal fusion based on dynamic weights to generate multi-source fused signals; ; Represents a multi-source fused signal; represent t Real-time bidding space signal; represent t Time-based price spread signals.

[0092] Furthermore, the weighting module includes: The random forest algorithm is used to select the top_p features with the highest importance weights. ; in, It is a tree tree All features used in f The set of nodes to be split, features f These are all feature columns except the target value; Impurity(s) It is a node s impurity, , , These are the parent node, left child node, and right child node, respectively. It is a tree tree Chinese characteristics f The weight.

[0093] Furthermore, the similar day module includes: Based on the bidding space signal of the prediction date, the data of the T dates closest to the bidding space signal are selected from the historical data through Pearson correlation coefficient calculation to construct a similar day data set for the prediction date, and the similar day data set is included in the training set.

[0094] The power trading risk monitoring system proposed in this embodiment, based on a multi-source signal mechanism and a sparse hybrid expert model, can achieve the power trading risk monitoring method based on a multi-source signal mechanism and a sparse hybrid expert model described in Embodiment 1, and has the same technical effect as Embodiment 1.

[0095] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A power trading risk monitoring method based on a multi-source signaling mechanism and a sparse hybrid expert model, characterized in that, include: S1. Collect historical day-ahead electricity market price data and raw data on influencing factors, perform data preprocessing, and construct a dataset; Based on the dataset, price spread signals and bidding space signals are constructed, and multi-source fusion signals are generated by combining dynamic weighted fusion algorithm and rolling window smoothing technology. S2. The random forest algorithm is used to calculate the importance weights of each influencing factor and the multi-source fusion signal on the price difference signal. Feature screening is performed, and the top_p features with the highest importance weights are selected to construct the training set and the test set. S3. Construct a similar day data set based on the bidding space signal, and incorporate the similar day data set into the training set; S4. The SparseMoE architecture of the sparse hybrid expert model is used as the model framework, and the feedforward neural network model is used as the basic expert model. Based on the training set data, different prediction models are trained with the price difference signal and the bidding space signal as the targets. S5. Based on the predicted price difference signal and bidding space signal obtained from the prediction model, a profit objective function is constructed. The assessment mechanism function is used as a penalty term, and the multi-source fusion signal is used as a condition. The original short-term power prediction is used as the strategy input parameter. The optimal application curve for the prediction day is dynamically searched through the Northern Eagle optimization search algorithm. Risk monitoring is achieved by comparing the consistency between the user application curve and the optimal application curve.

2. The power trading risk monitoring method based on multi-source signaling mechanism and sparse hybrid expert model according to claim 1, characterized in that, The preprocessing methods for the raw data of influencing factors in step S1 include: S101. Use the density-based DBSCAN clustering algorithm to perform outlier detection and analysis, and set outliers to null. S102. A three-stage hybrid filling strategy is adopted to fill null values ​​in the data.

3. The power trading risk monitoring method based on multi-source signaling mechanism and sparse hybrid expert model according to claim 1, characterized in that, The method for generating the multi-source fused signal in step S1 includes: S111, Constructing price spread signals and bidding space signals; S112. The correlation coefficient reflects the synergistic relationship between the price spread signal and the bidding space signal. Calculate the benchmark weights; ;in Represents the benchmark weight; This represents the weight sensitivity coefficient, with a value range of [0.1, 0.3]. S113. Calculate the dynamic weight of the bidding space signal using the benchmark weight; ； represent t Dynamic weighting of the bidding space signal at any given time; For bidding space signals in the window L within the standard deviation, For bidding space signals in the window L The mean within; S114. Signal fusion is performed based on dynamic weights to generate multi-source fused signals; ； Represents a multi-source fused signal; represent t Real-time bidding space signal; represent t Time-based price spread signals.

4. The power trading risk monitoring method based on multi-source signaling mechanism and sparse hybrid expert model according to claim 1, characterized in that, Step S2 includes: The random forest algorithm is used to select the top_p features with the highest importance weights. ； in, It is a tree tree All features used in f The set of nodes to be split, features f These are all feature columns except the target value; Impurity(s) It is a node s impurity, , , These are the parent node, left child node, and right child node, respectively. It is a tree tree Chinese characteristics f The weight.

5. The power trading risk monitoring method based on multi-source signaling mechanism and sparse hybrid expert model according to claim 1, characterized in that, Step S3 includes: Based on the bidding space signal of the prediction date, the data of the T dates closest to the bidding space signal are selected from the historical data through Pearson correlation coefficient calculation to construct a similar day data set for the prediction date, and the similar day data set is included in the training set.

6. A power trading risk monitoring system based on a multi-source signaling mechanism and a sparse hybrid expert model, characterized in that, include: Preprocessing and multi-source fusion signal module: Collects historical day-ahead electricity market price data and raw data of influencing factors, performs data preprocessing, and constructs datasets; Based on the dataset, price spread signals and bidding space signals are constructed, and multi-source fusion signals are generated by combining dynamic weighted fusion algorithm and rolling window smoothing technology. Weighting module: The random forest algorithm is used to calculate the importance weight of each influencing factor and the multi-source fusion signal on the price difference signal. Feature selection is performed to select the top_p features with the highest importance weights to construct the training set and test set. Similar Day Module: Construct a similar day dataset based on the bidding space signal, and incorporate the similar day dataset into the training set; Prediction Model Module: The SparseMoE architecture of the sparse hybrid expert model is used as the model framework, and the feedforward neural network model is used as the basic expert model. Based on the training set data, different prediction models are trained with the price spread signal and the bidding space signal as the targets. Risk monitoring module: Based on the price difference signal and bidding space signal obtained from the prediction model, a profit objective function is constructed. The assessment mechanism function is used as a penalty term, and multi-source fusion signals are used as conditions. The original short-term power prediction is used as the strategy input parameter. The optimal application curve for the prediction day is dynamically searched through the Northern Eagle optimization search algorithm. Risk monitoring is achieved by comparing the consistency between the user's application curve and the optimal application curve.

7. The power trading risk monitoring system based on a multi-source signal mechanism and a sparse hybrid expert model according to claim 6, characterized in that, The preprocessing and multi-source fusion signal module includes: Outlier handling unit: Employs density-based DBSCAN clustering algorithm for outlier detection and analysis, and sets outliers to null. Null value filling unit: A three-stage hybrid filling strategy is used to fill null values ​​in the data.

8. The power trading risk monitoring system based on a multi-source signal mechanism and a sparse hybrid expert model according to claim 6, characterized in that, The preprocessing and multi-source fusion signal module includes: Signal construction unit: Constructs price spread signals and bidding space signals; Benchmark weighting unit: The correlation coefficient reflects the synergistic relationship between the price spread signal and the bidding space signal. Calculate the benchmark weights; ;in Represents the benchmark weight; This represents the weight sensitivity coefficient, with a value range of [0.1, 0.3]. Dynamic weighting unit: Calculates the dynamic weight of the bidding space signal using the benchmark weight; ； represent t Dynamic weighting of the bidding space signal at any given time; For bidding space signals in the window L within the standard deviation, For bidding space signals in the window L The mean within; Fusion Unit: Performs signal fusion based on dynamic weights to generate multi-source fused signals; ； Represents a multi-source fused signal; represent t Real-time bidding space signal; represent t Time-based price spread signals.

9. The power trading risk monitoring system based on a multi-source signal mechanism and a sparse hybrid expert model according to claim 6, characterized in that, The weight module includes: The random forest algorithm is used to select the top_p features with the highest importance weights. ； in, It is a tree tree All features used in f The set of nodes to be split, features f These are all feature columns except the target value; Impurity(s) It is a node s impurity, , , These are the parent node, left child node, and right child node, respectively. It is a tree tree Chinese characteristics f The weight.

10. The power trading risk monitoring system based on a multi-source signal mechanism and a sparse hybrid expert model according to claim 6, characterized in that, The similar date module includes: Based on the bidding space signal of the prediction date, the data of the T dates closest to the bidding space signal are selected from the historical data through Pearson correlation coefficient calculation to construct a similar day data set for the prediction date, and the similar day data set is included in the training set.

Images